ChartOCR

Copy from original repo with the corner layers rewritten in pure PyTorch.

Getting Started

Create an Anaconda environment using the provided package list.

 conda create  --name chartocr --file conda_packagelist.txt

After you create the environment, activate it.

source activate chartocr

Inference can be run without a GPU.

Compiling NMS

You also need to compile the NMS code (originally from Faster R-CNN and Soft-NMS).

cd <CornerNet dir>/external
make

Installing MS COCO APIs

You also need to install the MS COCO APIs.

cd <CornerNet dir>/data
git clone git@github.com:cocodataset/cocoapi.git coco
cd <CornerNet dir>/data/coco/PythonAPI
make

Downloading CHARTEX Data

Pie data
Line data
Bar data
Cls data
Unzip the file to the data path

Data Description (Updated on 11/21/2021)

For Pie data
{"image_id": 74999, "category_id": 0, "bbox": [135.0, 60.0, 132.0, 60.0, 134.0, 130.0], "area": 105.02630551355209, "id": 433872}
The meaning of the bbox is [center_x, center_y, edge_1_x, edge_1_y, edge_2_x, edge_2_y]
It’s the three critical points for a sector of the pie graph.
For the line data
{"image_id": 120596, "category_id": 0, "bbox": [137.0, 131.0, 174.0, 113.0, 210.0, 80.0, 247.0, 85.0], "area": 0, "id": 288282}
The meaning of the bbox is [d_1_x, d_1_y, …., d_n_x,d_n_y]
It’s the data points for a line in the image with image_id.
instancesLineClsEx is used for training the LineCls.
For the Bar data
Just the bounding box of the bars.
For the cls data
Just the bounding box.
But different category_id refers to different components like the draw area, title and legends.

OCR API (Updated on 08/17/2022)

I am longger working at the microsoft, many features rely on the webservice may be out of date. The origninal OCR API requests the AZURE service. For people who do not have the AZURE service, pytesseract python pacakge may be a good replacment. However, you need to rewrite ocr_result(image_path) funtion. The key output of this function is the bounding box of the words and the str version of the words. E.g., word_info["text"]='Hello', word_info["boundingBox"] = [1, 2, 67, 78] The boudningBox is the topleft_x, topleft_y, bottomleft_x, bottomlef_y.

Downloading Trained File

data link
Unzip the file to current root path

Training and Evaluation

To train and evaluate a network, you will need to create a configuration file, which defines the hyperparameters, and a model file, which defines the network architecture. The configuration file should be in JSON format and placed in config/. Each configuration file should have a corresponding model file in models/. i.e. If there is a <model>.json in config/, there should be a <model>.py in models/. There is only one exception which we will mention later. The cfg file names of our proposed modules are as follows:

Bar: CornerNetPureBar

Pie: CornerNetPurePie

Line: CornerNetLine

Query: CornerNetLineClsReal

Cls: CornerNetCls

To train a model:

python train.py --cfg_file <model> --data_dir <data path> 
e.g. 
python train_chart.py --cfg_file CornerNetBar --data_dir /home/data/bardata(1031)

To use the trained model as a web server pipeline:

python manage.py runserver 8800

Access localhost:8800 to interact. This works very well and can be used to reference how the tool might be used for running inference only.

If you want to test batch of data directly, here you have to pre-assign the type of charts.

python test_pipe_type_cloud.py --image_path <image_path> --save_path <save_path> --type <type>
e.g.
python test_pipe_type_cloud.py --image_path /data/bar_test --save_path save --type Bar

fabianandresgrob/ChartOCR